aiqbench analysis
This research presents a training-free method to correct "shadow alignment" in large language models by fusing knowledge from well-aligned sources into adapted downstream tasks. By leveraging delta debugging and model fusion without requiring supervised fine-tuning or reinforcement learning, the approach significantly boosts defense success rates against harmful prompts while maintaining performance integrity. This work advances AI safety protocols for dynamic model adaptation scenarios where traditional retraining is impractical.
Decoupled Alignment for Robust Plug-and-Play Adaptation
arXiv:2406.01514v5 Announce Type: replace-cross Abstract: We introduce a training-free safety enhancement method for aligning large language models (LLMs) without the need for supervised fine-tuning or reinforcement learning from human feedback. Our main idea is to provide a robust plug-and-play approach to prevent shadow alignment when models are adapted to downstream tasks. Specifically, we leverage knowledge distillation to extract alignment signals from well-aligned LLMs and inject them into shadow-aligned models via model fusion, enabling plug-and-play alignment correction. In our methodology, we employ delta debugging to identify the critical components of knowledge necessary for effective distillation. On the harmful question dataset, our method significantly enhances the average defense success rate by approximately 14.42%, reaching as high as 51.39% across 17 influenced LLMs, without compromising performance. Our code is available at https://github.com/NWULIST/DAPA.
Abstract:We introduce a training-free safety enhancement method for aligning large language models (LLMs) without the need for supervised fine-tuning or reinforcement learning from human feedback. Our main idea is to provide a robust plug-and-play approach to prevent shadow alignment when models are adapted to downstream tasks. Specifically, we leverage knowledge distillation to extract alignment signals from well-aligned LLMs and inject them into shadow-aligned models via model fusion, enabling plug-and-play alignment correction. In our methodology, we employ delta debugging to identify the critical components of knowledge necessary for effective distillation. On the harmful question dataset, our method significantly enhances the average defense success rate by approximately 14.42%, reaching as high as 51.39% across 17 influenced LLMs, without compromising performance. Our code is available at this https URL.
Revised to correct the Acknowledgments section. Previous versions inadvertently included acknowledgments of NSF and NIH awards that did not support this work. Those funding acknowledgments have been removed. The technical content, results, and conclusions are unchanged
Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Cryptography and Security (cs.CR)
arXiv:2406.01514 [cs.CL]
(or arXiv:2406.01514v5 [cs.CL] for this version)
Full text under the publisher's open licence · 181 words
Related coverage
News
Quantum Computing Report · Jul 25, 2026
MemSFT: Mitigating Alignment Tax with an External Parametric Memory
arXiv · Jul 29, 2026
Open Rubric System: Scaling Reinforcement Learning with Pairwise Adaptive Rubric
arXiv · Aug 26, 2026
Rectifying Geometric Misalignment: Online Source-Free Adaptation for Class-Imbalanced EEG
arXiv · Aug 7, 2026